Module math 

Pure dense/sparse math helpers (testable without ORT).

Structs

SeqLenDistribution 🔒

Per-batch token-length distribution statistics.

Constants

SPECIAL_TOKENS 🔒

CLS, PAD, SEP/EOS, UNK — excluded from sparse output.

Functions

median_usize 🔒

Computes the median of a Vec<usize> in-place (sorts the slice).

normalize_l2 🔒

L2-normalizes vec in place. If the norm is zero, leaves the vector unchanged.

seq_len_distribution 🔒

Computes min, max, mean, and p95 token-length statistics for an embed batch.

sparse_maxpool 🔒

Max-pools sparse scores by vocabulary token ID, excluding special tokens and tokens masked by the attention mask.

sparse_project 🔒

Projects a single token’s hidden state through the sparse-linear layer.